One-Class Training for Masquerade Detection
Ke Wang, Salvatore J. Stolfo · 2003
We extend prior research on masquerade detection using UNIX commands issued by users as the audit source. Previous studies using multi-class training requires gathering data from multiple users to train specific profiles of self and non-self for each user. One-class training uses data representative of only one user. We apply one-class Naïve Bayes using both the multivariate Bernoulli model and the Multinomial model, and the one class SVM algorithm. The result shows that one-class training for this task works as well as multi-class training, with the great practical advantages of collecting much less data and more efficient training. One-class SVM using binary features performs best among the one-class training algorithms.